Dell High Performance Computing Solution Resources Owner's manual

Type
Owner's manual
Garima Kochhar
Dell HPC Engineering
March 2013, Version 1.0
Improving NFS Performance on HPC
Clusters with Dell Fluid Cache for DAS
This Dell technical white paper explains how to improve Network File
System I/O performance by using Dell Fluid Cache for Direct Attached
Storage in a High Performance Computing Cluster.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
2
This document is for informational purposes only and may contain typographical errors and
technical inaccuracies. The content is provided as is, without express or implied warranties of any
kind.
© 2013 Dell Inc. All rights reserved. Dell and its affiliates cannot be responsible for errors or omissions
in typography or photography. Dell, the Dell logo, PowerVault, and PowerEdge are trademarks of Dell
Inc. Intel and Xeon are registered trademarks of Intel Corporation in the U.S. and other countries.
Microsoft, Windows, and Windows Server are either trademarks or registered trademarks of Microsoft
Corporation in the United States and/or other countries. Other trademarks and trade names may be
used in this document to refer to either the entities claiming the marks and names or their products.
Dell disclaims proprietary interest in the marks and names of others.
March 2013| Version 1.0
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
3
Contents
Executive Summary ....................................................................................................... 5
1. Introduction ........................................................................................................... 6
1.1. Dell Fluid Cache for DAS (direct-attached storage) ..................................................... 6
2. Solution design and architecture.................................................................................. 6
2.1. NFS storage solution (baseline) ............................................................................. 7
2.2. Dell Fluid Cache for DAS based solution .................................................................. 9
2.3. I/O clients test bed ......................................................................................... 10
2.4. Solution tuning ............................................................................................... 12
2.4.1. Storage .................................................................................................. 13
2.4.2. NFS server .............................................................................................. 13
2.4.3. Dell Fluid Cache for DAS ............................................................................. 14
2.5. Tracking solution health and performance ............................................................. 14
2.5.1. Server health and monitoring ....................................................................... 14
2.5.2. Dell PowerEdge Express Flash PCIe SSD health and monitoring .............................. 14
2.5.3. Dell Fluid Cache for DAS health and monitoring ................................................. 16
3. Performance ........................................................................................................ 16
3.1. Sequential writes and reads ............................................................................... 17
3.2. Random writes and reads .................................................................................. 19
3.3. Metadata tests ............................................................................................... 20
3.4. Cold-cache tests ............................................................................................. 22
4. Conclusion ........................................................................................................... 25
5. References .......................................................................................................... 25
Appendix A: Step-by-step configuration of Dell Fluid Cache for NFS .......................................... 27
A.1. Hardware checklist and cabling .......................................................................... 27
A.2. NFS server set up ............................................................................................ 28
A.3. NFS server configuration and tuning ..................................................................... 29
A.4. Virtual disk configuration .................................................................................. 31
A.5. XFS and DFC configuration................................................................................. 33
A.6. Useful commands and references ........................................................................ 34
A.7. Performance tuning on clients ............................................................................ 34
Appendix B: Benchmarks and tests ................................................................................... 35
B.1. IOzone ......................................................................................................... 35
B.2. mdtest ......................................................................................................... 37
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
4
Tables
Table 1. NFS server and storage hardware configuration ...................................................... 8
Table 2. NFS server software and firmware configuration ..................................................... 9
Table 3. Hardware configuration for DFC ....................................................................... 10
Table 4. Software and firmware configuration for DFC ...................................................... 10
Table 5. I/O cluster details ........................................................................................ 11
Figures
Figure 1. NFS storage solution ....................................................................................... 7
Figure 2. NFS server ................................................................................................... 8
Figure 3. Test bed ................................................................................................... 12
Figure 4. PCIe SSD health ........................................................................................... 15
Figure 5. Large sequential write performance.................................................................. 18
Figure 6. Large sequential read performance .................................................................. 18
Figure 7. Random write performance ............................................................................ 19
Figure 8. Random read performance ............................................................................. 20
Figure 9. Metadata file create performance .................................................................... 21
Figure 10. Metadata file stat performance ....................................................................... 21
Figure 11. Metadata file remove performance ................................................................... 22
Figure 12. Cold-cache sequential reads ........................................................................... 23
Figure 13. Cold-cache random reads ............................................................................... 23
Figure 14. CacheIO-DiskIO on cold-cache reads .................................................................. 24
Figure 15. Solution cabling .......................................................................................... 28
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
5
Executive Summary
Most High Performance Computing clusters use some form of a Network File System (NFS) based storage
solution for user data. Easy to configure and administer, free with virtually all Linux distributions, and
well-tested and reliable, NFS has many advantages. Use of nearline SAS drives for backend storage
provides large capacity and good performance at a reasonable cost, but with an inherent performance
limitation for random I/O patterns.
This technical white paper describes how to improve I/O performance in such a NFS storage solution
with the use of Dell Fluid Cache for DAS (DFC) technology. It describes the solution and presents
cluster-level measured results for several I/O patterns. These results quantify the performance
improvements possible with DFC, especially for random I/O patterns. This white paper also includes a
how-to recipe in the Appendix that provides step-by-step instructions on building the solution.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
6
1. Introduction
A Network File System (NFS) based storage solution is a popular choice for High Performance
Computing Clusters (HPC). Most HPC clusters use some form of NFS irrespective of the size of the
cluster. NFS is simple to configure and administer, free with virtually all Linux distributions, well-
tested, and can provide reliable storage for user home directories and application data. A well-tuned
NFS solution can provide great performance for small to mid-sized clusters.
1
Nearline SAS (NL SAS)
drives provide large capacity and good performance for a reasonable price, optimizing the GB/$
metric. However, they limit performance for applications that have random I/O patterns.
2
Dell Fluid
Cache for DAS (Direct Attached Storage) reduces this limitation by caching data while the backend
storage services the I/O request, thus improving the performance of the entire NFS storage solution.
This study evaluates Dell Fluid Cache for DAS (DFC)
3
with NFS for HPC clusters. A cluster-level study
was conducted to analyze different I/O characteristics and quantify the performance improvements
gained with DFC when compared to the same NFS configuration without using DFC. All the results
presented in this document are measured results that were obtained in the Dell HPC laboratory.
The following section introduces the DFC technology. Subsequent sections describe the design of the
storage solution and the tuning optimizations applied to the solution. Information is provided on tools
that can be used to monitor the solution. An analysis of the performance results follows. The paper
concludes with recommendations on best-fit use cases for DFC with NFS.
Two appendices that provide step-by-step instructions on how to configure such a solution and provide
information on the benchmarks and tests that were run for this study complete this document.
1.1. Dell Fluid Cache for DAS (direct-attached storage)
DFC is a write-back host caching software. DFC combines multiple Dell PowerEdgeExpress Flash PCIe
SSDs to provide a read and write cache pool. This PCIe SSD cache pool is used to accelerate response
times with significant improvements in I/O operations per second (IOPS).
Some features of the DFC software include:
Faster cache reads, writes, read-after-writes, and re-reads
Data protection as writes are replicated across multiple PCIe SSDs
Orderly hot swap and hot plug capability that allows adding or removing a device without
halting or rebooting the system
More details on the Dell Fluid Cache for DAS technology can be found at [3].
In an HPC context, the DFC software can be configured on a NFS server. PCIe SSDs on the NFS server
will provide the virtual cache pool.
2. Solution design and architecture
This section describes the NFS storage solution used to evaluate the DFC technology. The baseline for
comparison is an NFS server with direct-attached external SAS storage. The configuration of this NFS
server is augmented with PCIe SSDs and DFC software for the DFC comparison. A 64-server Dell
PowerEdge cluster was used as I/O clients to provide I/O load to the storage solution. The following
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
7
sections provide details on each of these components as well as information on tuning and monitoring
the solution.
2.1. NFS storage solution (baseline)
The baseline in this study is an NFS configuration. One PowerEdge R720 is used as the NFS server.
PowerVaultMD1200 storage arrays are direct-attached to the PowerEdge R720 and provide the
storage. The attached storage is formatted as a Red Hat Scalable File System (XFS). This file system is
exported via NFS to the HPC compute cluster.
The NFS server and storage is shown in Figure 1 and Figure 2. Table 1 and Table 2 list the details of the
configuration. Readers familiar with Dell’s NSS line of solutions
1
will recognize this baseline
configuration.
NFS storage solution Figure 1.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
8
NFS server Figure 2.
NFS server and storage hardware configuration Table 1.
Server configuration
NFS SERVER PowerEdge R720
PROCESSORS Dual Intel(R) Xeon(R) CPU E5-2680 @ 2.70 GHz
MEMORY 128 GB. 16 * 8 GB 1600MT/s RDIMMs
INTERNAL DISKS
5 * 300 GB 15 K SAS disks
Two drives configured in RAID-0 for the operating system
with one additional drive as a hot spare
Two drives configured in RAID-1 for swap space
INTERNAL RAID CONTROLLER PERC H710P mini (internal)
EXTERNAL RAID CONTROLLER
PERC H810 adapter (slot 7) connected to the storage
enclosures
INTERCONNECT TO CLIENTS Mellanox ConnectX-3 FDR card (slot 5)
Storage configuration
Storage Enclosure Four PowerVault MD1200 arrays, daisy chained
Hard Disks 12 * 3 TB 7200 rpm NL SAS drives per storage enclosure
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
9
NFS server software and firmware configuration Table 2.
Software
OPERATING SYSTEM Red Hat Enterprise Linux (RHEL) 6.3z
KERNEL VERSION 2.6.32-279.14.1.el6.x86_64
FILE SYSTEM Red Hat Scalable File System (XFS) 3.1.1-7
SYSTEMS MANAGEMENT Dell OpenManage Server Administrator 7.1.2
Firmware and Drivers
BIOS 1.3.6
iDRAC 1.23.23 (Build 1)
PERC H710/PERC H810
FIRMWARE
21.1.0-0007
PERC DRIVER megasas 00.00.06.14-rh1
INFINIBAND FIRMWARE 2.11.500
INFINIBAND DRIVER Mellanox OFED 1.5.3-3.1.0
The baseline described in this section is very similar to the Dell NSS. One key difference is the use of a
single RAID controller to connect to all four storage arrays. In a pure-NSS environment, two PERC RAID
controllers are recommended for optimal performance. With two PERC cards, the two RAID virtual disks
are combined using Linux Logical Volume Manager (LVM). DFC does not support caching of an LVM
device, hence a single PERC was used for this study.
The NFS server and the attached storage arrays are configured and tuned for optimal performance
based on several past studies
4
. A summary of the design choices is provided in Section 2.4.
Detailed instructions on configuring this storage solution are provided in Appendix A: Step-by-step
configuration of Dell Fluid Cache for NFS.
2.2. Dell Fluid Cache for DAS based solution
The DFC-based NFS solution builds on top of the baseline configuration described in Section 2.1. It
simply adds PCIe SSDs and the DFC software to the baseline configuration. Details of the configuration
are provided in Table 3 and Table 4.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
10
Hardware configuration for DFC Table 3.
Server configuration
NFS SERVER PowerEdge R720
CACHE POOL Two 350GB Dell PowerEdge Express Flash PCIe SSD
SSD CONTROLLER Internal (slot 4)
Rest of the configuration is the same as baseline, as described in Table 1
Storage configuration
Same as baseline, as described in Table 1
Software and firmware configuration for DFC Table 4.
Software
CACHING SOFTWARE Dell Fluid Cache for DAS v1.0
Rest of the configuration is the same as baseline, as described in Table 2
Firmware and Drivers
PCIe SSD DRIVER
mtip32xx 1.3.7-1 latest available at the time of this study.
Recommend using mtip32xx v2.1.0
Rest of the configuration is the same as baseline, as described in Table 2
In DFC vocabulary, the cache or cache pool is the SSDs, and the disk that is enabled for caching is the
virtual disk on the PowerVault MD1200s. Most importantly, the methods used to access the data remain
the same as in the baseline case. The I/O clients simply mount the same NFS exported directory as in
the baseline configuration. Detailed instructions on configuring DFC for this storage solution are
provided in Appendix A: Step-by-step configuration of Dell Fluid Cache for NFS.
2.3. I/O clients test bed
The pure NFS baseline solution and the NFS+DFC solution were exercised using a 64-node HPC cluster.
This compute cluster was used to provide I/O load to the storage solution and help benchmark the
capabilities of the solution.
Using the latest quarter height Dell PowerEdge M420 blade server
5
as the building block for the I/O
cluster, the 64-client cluster was configured in 20U of rack space. Details of the 64-client test bed are
provided in Table 5. Figure 3 shows the entire test bed including the clients. Note that all I/O traffic to
the NFS server used the InfiniBand network and the IPoIB protocol.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
11
I/O cluster details Table 5.
I/O cluster configuration
CLIENTS
64 PowerEdge M420 blade servers
32 blades in each of two PowerEdge M1000e chassis
CHASSIS CONFIGURATION
Two PowerEdge M1000e chassis, each with 32 blades
Two Mellanox M4001F FDR10 I/O modules per chassis
Two PowerConnect M6220 I/O switch modules per chassis
INFINIBAND FABRIC
For I/O traffic
Each PowerEdge M1000e chassis has two Mellanox M4001
FDR10 I/O module switches.
Each FDR10 I/O module has four uplinks to a rack Mellanox
SX6025 FDR switch for a total of 16 uplinks.
The FDR rack switch has a single FDR link to the NFS server.
ETHERNET FABRIC
For cluster deployment and
management
Each PowerEdge M1000e chassis has two PowerConnect
M6220 Ethernet switch modules.
Each M6220 switch module has one link to a rack
PowerConnect 5224 switch.
There is one link from the rack PowerConnect switch to an
Ethernet interface on the cluster master node.
I/O compute node configuration
CLIENT PowerEdge M420 blade server
PROCESSORS Dual Intel(R) Xeon(R) CPU E5-2470 @ 2.30 GHz
MEMORY 48 GB. 6 * 8 GB 1600 MT/s RDIMMs
INTERNAL DISK 1 50GB SATA SSD
INTERNAL RAID CONTROLLER PERC H310 Embedded
CLUSTER ADMINISTRATION
INTERCONNECT
Broadcom NetXtreme II BCM57810
I/O INTERCONNECT
Mellanox ConnectX-3 FDR10 mezzanine card
I/O cluster software and firmware
BIOS 1.3.5
iDRAC 1.23.23 (Build 1)
OPERATING SYSTEM Red Hat Enterprise Linux (RHEL) 6.2
KERNEL 2.6.32-220.el6.x86_64
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
12
OFED Mellanox OFED 1.5.3-3.0.0
Test bed Figure 3.
2.4. Solution tuning
The NFS server and the attached storage arrays are configured and tuned for optimal performance.
These options were selected based on extensive studies done by the Dell HPC team. Results of these
studies and the tradeoffs of the tuning options are available in [4].
Additionally the DFC configuration was tuned based on experience gained from this study.
This section provides a quick summary of some of the optimizations applied to the storage solution.
Detailed instructions on configuring this storage solution are provided in Appendix A: Step-by-step
configuration of Dell Fluid Cache for NFS.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
13
2.4.1. Storage
3TB NL SAS disks are selected for large capacity at a cost-effective price point.
Virtual disks are created using a RAID 60 layout. The RAID 6 span is across 10 data disks and 2
parity disks and the stripe is across all four storage enclosures. This RAID configuration provides
a good balance between capacity, reliability to tolerate multiple disk failures and
performance.
4
The segment size used for the RAID stripe is 512 KB to maximize performance.
4
This value
should be set based on the expected application I/O profile for the cluster.
2.4.2. NFS server
The default OS scheduler is changed from cfq to deadline to maximize I/O performance.
4
The number of concurrent nfsd threads is increased to 256 from the default of 8 to maximize
performance.
4
RHEL 6.3 errata Kernel version 2.6.32-279.14.1 fixes some important XFS bugs
6
and is
recommended for this solution.
The PowerEdge R720 is a dual-socket system that uses the Intel Xeon E5-2600 series processors.
On these processors, the PCI-E controller is integrated on the processor chip, making some slots
‘closer’ to one socket and further from the other. This makes card-to-PCI slot mapping an
important factor to consider in performance tuning. The solution presented in this white paper
balances the three PCI cards across the two processors based on server design and engineering
best practices. Refer to Table 1 and Table 3 for card-to-slot recommendations.
Two internal disks on the NFS server are configured in a RAID 0 stripe and used as swap space
for the operating system. This provides a large swap space in case there is a need for the XFS
repair program to run after an ungraceful system shutdown.
XFS create options are optimized as well. By default, XFS tries to query the underlying storage
device and optimize the settings accordingly. In the case of using LVM, this works fine;
however, when presenting a raw virtual disk to XFS, it is important to specify the stripe unit
(su) and stripe width (sw). The stripe unit is the stripe element size that was used to format
the virtual disk. The stripe width is the number of data drives in the virtual disk.
By default, there are 8 log buffers, and the -l size option tells xfs how large the log buffers can
become. This can improve metadata performance; however, the larger the log, the longer it
may take to recover a file system that was not unmounted cleanly.
In this solution the mkfs.xfs command used was:
mkfs.xfs -d su=512k,sw=40 -l size=128m /dev/sdc
There are XFS mount options to optimize the file system as well. The options used for this
solution are similar to the Dell NSS and are noatime, allocsize=1g, nobarrier, inode64,
logbsize=262144, attr2. Details of these mount options are provided in [4].
The “sync” NFS export option is used when exporting the XFS file system at the expense of
lower performance. This is an added layer of protection to ensure data reliability as the NFS
server will not acknowledge the write until it has written the data to disk.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
14
NFSv3 is recommended over NFSv4 based on the performance results of a previous study.
4
It
was found that metadata create operations have significantly lower performance when using
NFSv4. For environments where the security enhancements in NFSv4 are more important than
performance considerations, NFSv4 can be used instead.
2.4.3. Dell Fluid Cache for DAS
In tests conducted at the Dell HPC Engineering laboratory, it was found that enabling the
“Logical Processor” option in the NFS server BIOS improves the random read results of the DFC
configuration by up to 48% when compared to “Logical Processor” Disabled. Dell recommends
enabling the Logical Processor setting when using DFC.
All results in this document have been achieved with this setting enabled.
DFC can support one to four SSDs in the solution. The DFC cache pool configuration used in this
solution uses two PCIe SSDS. Two SSDs provide ~650GB for caching of reads. Two SSDS are the
minimum required to configure DFC in write-back mode, where writes are cached and
replicated, thus helping to accelerate writes as well as re-reads.
3
Caching capacity can be further increased by using four SSDSs in the solution. This will enable
the NFS server to sustain top-line performance for larger amounts of data and/or to support a
larger number of simultaneous clients. Some I/O patterns presented in Section 3 such as large
sequential writes will see improved performance with a larger cache pool.
2.5. Tracking solution health and performance
From a system administrator’s point of view, there are several components in this storage solution that
need to be monitored. This section lists a few simple Dell utilities that can be used to track the
solution’s health and performance statistics.
2.5.1. Server health and monitoring
Dell OpenManage Server Administrator (OMSA)
7
is a Dell utility used to administer and monitor Dell
PowerEdge servers. For the purpose of this solution, OMSA can be used to configure the system BIOS,
create virtual disks both on internal and external drives and monitor the health of the system.
OMSA provides a command-line interface as well as graphical user interface that can be used for all
systems management tasks.
Dell recommends that OMSA’s server and storage management components be installed on the NFS
server in this solution. OMSA v7.1.2 provides support for DFC.
2.5.2. Dell PowerEdge Express Flash PCIe SSD health and monitoring
Configuring the PCIe SSDs in the NFS server is a very straightforward task. The drivers are controlled by
the internal SSD controller. Dell recommends installing the latest Dell drivers to interface with the
SSDs. The version used in this solution is provided in Table 4.
The Express Flash PCIe SSDs are an enterprise class SSD. The performance of the SSD is guaranteed for
the lifetime of the device within the warranty period. There is no expectation of any performance
degradation as the device ages, and blocks are written to the device within the warranty lifetime of
the device.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
15
The warranty of the device is expressed in number of years and number of Petabytes written (PBW).
For the recommended 350GB SSD drive, the standard warranty is 3 years, 25 PBW.
The health of the device can be monitored using Dell OMSA utilities. OMSA reports the SSD “Device Life
Remaining” and “Failure Predicted”. “Device Life Remaining” is an indication of the amount of data
written to the device, and is calibrated for the PBW portion of the warranty. It is tracked internally by
the SSD. “Failure Predicted” tracks the health of the internal components of the SSD. If failure is
predicted, it is recommended that the SSD be replaced. This field does not imply performance
degradation but suggests that the SSD is starting to have component failure counts that are higher than
the preset thresholds.
Example output is shown in Figure 4. For more details refer to [8].
PCIe SSD health Figure 4.
[root@nfs-dfc~]# omreport storage controller controller=3
Controller PCIe-SSD SubSystem (Not Available)
Controllers
ID : 3
Status : Ok
Name : PCIe-SSD SubSystem
<…snip…>
PCIe-SSD Extender
ID : 0
Status : Ok
Name : PCIe-SSD Extender
State : Ready
Connector Type : Non-RAID
Termination : Not Applicable
SCSI Rate : Not Applicable
Physical Disks
ID : 0:2:0
Status : Ok
Name : Physical Device 0:2:0
State : Ready
Power Status : Not Applicable
Bus Protocol : PCIe
Media : SSD
Device Life Remaining : 100%
Failure Predicted : No
Revision : B1490208
Driver Version : 1.3.7
Model Number : DELL_P320h-MTFDGAL350SAH
<…snip…>
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
16
2.5.3. Dell Fluid Cache for DAS health and monitoring
DFC provides a very simple command-line utility /opt/dell/fluidcache/bin/fldc that can be
used for configuration and management. Alternately, the DFC configuration can be accomplished using
the OMSA GUI. DFC is a component under the storage sub-section of the OMSA GUI.
/opt/dell/fluidcache/bin/fldcstat is a command-line utility that provides extensive statistics
of the cache hits on the system, disk IO, etc.
Additional details are available in the DFC User’s Guide in [3].
3. Performance
This section presents the results of performance tests conducted on the storage solution described in
Section 2. Performance tests were run to evaluate the following common I/O patterns.
Large sequential reads and writes
Small random reads and writes
Metadata operations
These tests were performed over the IP-over-InfiniBand (IPoIB) network as described in Section 2.3.
The iozone and mdtest benchmarks were used for this study. Details of the benchmarks and test
process are provided in Appendix B: Benchmarks and tests.
Iozone was used for the sequential tests as well as the random tests. The I/O access patterns are N-to-
N, i.e., each thread reads and writes to its own file. Iozone was executed in clustered mode and one
thread was launched on each compute node. For the sequential tests, the performance metric used
was throughput in terms of MiB/s. For random tests, I/O operations per second (IOPS) was the metric.
The large sequential read and large sequential write tests were conducted using a request size of 1024
KB. The total amount of data written was 256 GB. (Recall from Table 1 that the NFS server RAM is 128
GB.) This is to ensure that the total I/O exceeds the NFS server memory since the goal is to test the
disk and storage solution performance.
The small random tests were performed with 4 KB record sizes since the size corresponds to typical
random I/O workloads. Each client reads and writes a smaller 4 GB file for these tests.
The metadata tests were performed with the mdtest utility and include file creates, stats, and
removals.
While these benchmarks do not cover every I/O pattern, they help characterize the I/O performance of
this storage solution.
Each set of tests was run on a range of clients to test the scalability of the solution. The number of
simultaneous clients involved in each test was varied from one to 64 clients. The client test bed is
described in Section 2.3.
Tests were performed on three configurations:
Baseline - This is the pure NFS storage solution as described in Section 2.1.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
17
DFC in Write-Back mode (DFC-WB) This configuration builds on the baseline by adding DFC as
described in Section 2.2, and DFC is configured to operate in Write-Back (WB) mode. WB mode
allows the caching of writes on the cache pool. WB mode requires the data to be written to a
minimum of two PCIe SSDs. Both re-reads and writes are accelerated.
DFC in Write-Through mode (DFC-WT) This configuration builds on the baseline, but here DFC
is configured in Write-Through (WT) mode. WT mode forces writes to both the cache and
virtual back end disk simultaneously.
The following sections present the results of the different I/O patterns.
3.1. Sequential writes and reads
The results of the IPoIB sequential write tests are shown in Figure 5. The figure shows the aggregate
throughput that can be achieved when a number of clients are simultaneously writing to the storage
over the InfiniBand fabric.
The results show that baseline configuration can reach a peak write throughout of ~2,000 MiB/s. Recall
that this is with the “sync” NFS export option, and this peak throughput demonstrates how well the
configuration is tuned
4
. With DFC in WB mode, labeled DFC-WB in Figure 5, the throughput measured
was ~400 MiB/s with a peak of ~600 MiB/s. The lower sequential write performance with DFC is due to
two factors: the pure sequential write performance of the SSDs is lower than the storage array and the
write-back cache has to replicate dirty blocks which means every write to the cache has to be written
to two SSDs. Subsequent re-write operations were found to have ~25% higher throughput as the files
are already in the DFC cache.
Switching to WT mode and eliminating the replica blocks doubles the sequential write performance.
Peak performance approaches ~1,000 MiB/s as seen in Figure 5, labeled DFC-WT. Recall that all writes
go directly to the backend disk in addition to being cached on the SSDs in WT mode.
If the I/O pattern is such that there is a large amount of sequential data written to the backend
storage initially followed by subsequent reads, re-reads, and small writes, one method to take
advantage of DFC performance on random workload (Section 3.2) while minimizing the sequential write
performance penalty is to disable caching on the backend disk during the write operation. Although this
might not be an option for production clusters, smaller single-user/single-application environments
might be able to adopt this approach. DFC provides very simple utilities to accomplish this. Once the
data is written to the backend store, for example by a gene sequencer, caching on the backend disk
can be re-enabled. Subsequent reads, re-reads, and writes can benefit greatly from DFC technology.
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
18
Large sequential write performance Figure 5.
Large sequential read performance Figure 6.
Figure 6 shows the results of the sequential read tests. The aggregate throughput achieved when a
number of clients are simultaneously reading files is plotted on the graph. The figures show the peak
read throughput for the baseline is ~2,500 MiB/s. With the DFC configuration, reads are 13% to 60%
0
500
1000
1500
2000
2500
1 2 4 8 16 32 48 64
Throughput in MiB/s
Number of concurrent clients
Sequential writes
baseline DFC-WB DFC-WT
0
500
1000
1500
2000
2500
3000
3500
1 2 4 8 16 32 48 64
Throughput in MiB/s
Number of concurrent clients
Sequential reads
baseline DFC-WB DFC-WT
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
19
better than the baseline since the data is already in the DFC cache. As expected on read operations,
WB and WT tests have similar performance and can reach peak throughout of ~3050 MiB/s.
3.2. Random writes and reads
Figure 7 plots the aggregate IOPs when a number of clients are simultaneously issuing random writes to
their files. The baseline configuration can sustain ~1,600 IOPs on writes. Random writes are limited by
capability of the RAID controller and the disks seek latency of the backend disks.
With DFC in WB mode, the true value of DFC becomes apparent. Write IOPs are up to 6.4 times higher
(6.4x) than the baseline configuration at 64 clients. In random write tests, the cache warms up very
quickly since random operations tend to be ‘small’ with a 4k block size. In WT mode, DFC behavior and
performance is similar to the baseline as all writes go through to the backend disk.
Figure 8 plots the aggregate read IOPs when a number of clients are simultaneously issuing random
read operations. The baseline configuration sustains ~9,300 IOPS. As expected on read operations, DFC
WB and WT tests have similar performance and peak at 123,000 IOPs at 32 clients! That is over 20x
higher than the baseline performance.
Clearly read requests are being fulfilled from the DFC cache and that is the value of such caching
software. The question that arises is about worst case read performance. What is the drop in
performance when the data is accessed after it has been evicted from the cache? This scenario is
explored in Section 3.4.
Random write performance Figure 7.
0
2000
4000
6000
8000
10000
12000
14000
1 2 4 8 16 32 48 64
IOPS
Number of concurrent clients
Random writes
baseline DFC-WB DFC-WT
Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
20
Random read performance Figure 8.
3.3. Metadata tests
This section presents the results of metadata tests using the mdtest benchmark. In separate tests, one
million files were created, stated and unlinked concurrently from multiple NFS clients on the NFS
server. These results are presented in Figure 9, Figure 10, and Figure 11, respectively, as the number
of metadata operations per second. Each client ran a single instance of the mdtest benchmark for test
cases from 1 to 64 clients. For 128, 256 and 512 concurrent client tests, each client ran 2, 4, and 8
threads of the benchmark.
From the figures it can be seen that DFC-WT and DFC-WB create and remove tests have similar
performance. This indicates that the metadata tests are not bandwidth sensitive. If they were, the WT
tests would be expected to outperform the WB tests. This result needs further analysis. It is likely that
these operations are more latency sensitive.
0
20000
40000
60000
80000
100000
120000
140000
1 2 4 8 16 32 48 64
IOPS
Number of concurrent clients
Random reads
baseline DFC-WB DFC-WT
  • Page 1 1
  • Page 2 2
  • Page 3 3
  • Page 4 4
  • Page 5 5
  • Page 6 6
  • Page 7 7
  • Page 8 8
  • Page 9 9
  • Page 10 10
  • Page 11 11
  • Page 12 12
  • Page 13 13
  • Page 14 14
  • Page 15 15
  • Page 16 16
  • Page 17 17
  • Page 18 18
  • Page 19 19
  • Page 20 20
  • Page 21 21
  • Page 22 22
  • Page 23 23
  • Page 24 24
  • Page 25 25
  • Page 26 26
  • Page 27 27
  • Page 28 28
  • Page 29 29
  • Page 30 30
  • Page 31 31
  • Page 32 32
  • Page 33 33
  • Page 34 34
  • Page 35 35
  • Page 36 36
  • Page 37 37
  • Page 38 38

Dell High Performance Computing Solution Resources Owner's manual

Type
Owner's manual

Ask a question and I''ll find the answer in the document

Finding information in a document is now easier with AI